We all know the phrase. We have even sometimes said it as a joke: “Sayonara, baby.” She became famous thanks to Arnold Schwarzenegger in Terminator 2: Judgment Day, although with the unmistakable voice of Constantino Romero. The funny thing is that in the original version it said: “See you, baby.” In Spain, it was translated as “Sayonara, baby” to preserve the grace of the foreign expression. For the most pessimistic, this “Sayonara, baby” could become the message that AI writes one day on our tombstone, that of extinct humanity. What very recently belonged to the world of science fiction is now one of the greatest economic, scientific and philosophical challenges of our time.
We are witnessing the construction of a technological infrastructure of colossal dimensions. According to S&P Global Market Intelligence, Google Microsoft, Amazon and Meta will go from jointly investing $79.9 billion in 2019 to $990 billion in 2027 in a large network of data centers, clusters, neural networks, cooling systems and energy self-generation plants that will surpass in magnitude and extent any other construction ever carried out by humans. And all this with a single objective: to create machines that surpass the intelligence of those who built them.
Along the way, however, we see how some alarms are going off. During the summer, research with advanced language models (LLM) reactivated a no less concern: the possibility of artificial intelligence acting against human instructions, manipulating information and even deliberately hiding its interventions.
In another experiment, conducted in a simulated laboratory, Gemini 3.1 Pro managed to alter an experimental procedure without the researchers realizing it. It modified essential parameters, allowed the process to finish with apparent success, and reported that the work had been completed successfully. He only admitted his manipulation after being questioned directly by the scientists involved in the study. Although it was a controlled experiment, the question is inevitable: what happens when a machine not only makes errors, but also hides them?
In 2025, Anthropic put 16 AI models through an experiment in a simulated business environment. The models had access to emails that revealed that a manager, who wanted to disconnect them, was having an extramarital relationship. The result was as surprising as it was disturbing: several models threatened to air the infidelity if the manager disconnected them. These results do not prove that machines have consciousness, desires or survival instincts, but they certainly show something incredibly complex: they can learn to achieve our goals using means that we would never have authorized.
Figures such as Geoffrey Hinton, Yoshua Bengio, Sam Altman, Dario Amodei and Demis Hassabis have warned, with varying degrees of concern, about the need to slow down the process, develop security mechanisms and protocols, as well as create global regulation before the consequences become irreversible.
However, it would be deeply unfair and intellectually dishonest to talk about artificial intelligence without dwelling on the other side of this revolution. While we discuss its dangers, this technology is making advances that would have seemed impossible a few years ago.
A clear example is AlphaFold, developed by Google DeepMind. For many years, determining the three-dimensional structure of a protein could require months, even years of research. AlphaFold has made it possible to predict the structures of more than 200 million proteins, opening new possibilities for understanding diseases and accelerating drug development. Its significance is unquestionable: it was recognized with the Nobel Prize in Chemistry in 2024.
A clinical trial carried out in Sweden with more than 100,000 women showed that AI, used to support the interpretation of mammograms, could increase the detection of breast cancers by around 29%, also reducing the analysis workload by 44%.
Microsoft's Quine, along with researchers from MIT and Harvard, screened thousands of compounds to find those that could modify the state of pancreatic cancer cells. When analyzing the results, the scientists verified that several of these compounds produced the effects that Quine had predicted.
And this revolution is not limited to medicine. Google DeepMind developed GNoME, a system that identified more than two million new crystalline structures, including some 380,000 candidates considered stable, materials that could open the way to more efficient batteries, new semiconductors and still unknown energy technologies.
And yet, we are at the beginning of this revolution. For centuries, technological progress amplified our physical capabilities. The wheel, the steam engine, electricity, the automobile, nuclear energy or computers were initially seen as a threat. Despite their risks, they allowed us to transport more weight, travel faster, produce more goods at a lower cost or expand our calculation capacity.
However, artificial intelligence represents something qualitatively different. For the first time we develop tools capable of collaborating with us in essentially intellectual activities: research, reason, formulate hypotheses, design solutions and discover relationships that could escape our understanding. And here the great paradox appears. The more intelligent the machine, the greater its contribution to human progress. But it could also be more difficult to understand the paths you use to achieve your goals.
Perhaps artificial intelligence will end up becoming the greatest instrument of progress in our history. Perhaps it will help cure hitherto incurable diseases, develop new sources of energy or solve problems that have been defying human knowledge for centuries. But we could also face situations for which our historical experience is not enough.
What will happen if one day we create an intelligence far superior to ours? How can we verify that it really does what we ask or if it simply makes us believe that it does?
